DGHR-YOLO: fabric defect detection based on High-level Screening -feature Pyramid Networks and deformable convolution

作     者：Zhang, Zhixing Huang, Tongyuan Zhang, Weifeng Yang, Yihan Yu, Qianjiang 

作者机构：Chongqing Univ Technol Sch Artificial Intelligence Chongqing Peoples R China 

出 版 物：《NONDESTRUCTIVE TESTING AND EVALUATION》 (Nondestr Test Eval)

年 卷 期：2025年

核心收录：

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 08[工学] 0805[工学-材料科学与工程（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China Chongqing Natural Science Foundation [CSTB2022NSCQ-MSX0933] Science and Technology Foundation of Chongqing Education Commission [KJQN202201144] 

主　　题：Convolutional neural network deformable convolution network GSConv High-level Screening-feature Pyramid Networks RCS-OSA Fabric defect detection 

摘      要：Fabric defect detection is crucial for quality control in fabric manufacturing but remains a challenge due to the multi-scale characteristics of defects and their integration with the fabric background. To address this, we propose an efficient fabric defect detection model, DGHR-YOLO. First, the deformable convolution block (DCB) is introduced into the backbone network, leveraging its dynamic receptive field to effectively capture defect morphology and enhance feature extraction. Secondly, we propose the GS-SPPF module, designed to mitigate semantic information loss, optimize feature fusion, and improve both accuracy and inference speed. Thirdly, a lightweight High-level Screening Feature Pyramid Network (HS-FPN) is introduced, enabling effective multi-scale feature fusion while maintaining low complexity. Finally, a one-shot aggregation module based on channel shuffle and re-parameterized convolution is introduced, enhancing feature interaction across scales. Experimental results on the Tianchi textile dataset demonstrate that DGHR-YOLO achieves mAP0.5 and mAP0.5:0.95 scores of 85.8% and 72.7%, with respective improvements of 2.7% and 3.8% over YOLOv8m, while maintaining low parameter count and computational complexity.